Soft Sensing for Time Series With Irregular Sampling Internals Based on a Denoising Interval Attention LSTM Network

Area of Science:

• Chemical Engineering
• Data Science
• Industrial Process Control

Background:

• Industrial monitoring relies on predicting key quality variables.
• Data acquisition challenges include high noise and irregular sampling.
• Existing methods struggle with these data imperfections.

Purpose of the Study:

• To develop an advanced soft sensing model for industrial applications.
• To address challenges of noisy and irregularly sampled data.
• To improve the accuracy of predicting key quality variables.

Main Methods:

• A stacked supervised and reconstructed input denoising autoencoder (SSRDAE) was designed.
• The SSRDAE extracts quality-related features while minimizing information loss.
• An interval attention long short-term memory (IALSTM) network processed denoised features to capture temporal dependencies.

Main Results:

• The SSRDAE-IALSTM model demonstrated enhanced learning of process features.
• Superior prediction performance was achieved compared to existing methods.
• Validation on a debutanizer column and penicillin fermentation confirmed effectiveness.

Conclusions:

• The proposed SSRDAE-IALSTM network offers a robust solution for soft sensing under challenging industrial data conditions.
• The model effectively integrates feature extraction and temporal modeling for accurate quality prediction.
• This approach advances industrial status identification and monitoring capabilities.